TabPFN-4: The next step for predictive intelligence
November 19, 2026
•
10 min
Our next-generation tabular foundation model, TabPFN-4, has finished training. It breaks the quadratic barrier and moves the Pareto frontier on all evaluated benchmarks. Its novel sub-quadratic architecture scales to 10 million rows, while setting new Pareto frontiers in BeyondArena and TabArena.
TabPFN-4 will be available soon. Join the waitlist to get access first:
RelArena-α: the TabPFN columns run the TabPFN-Rel recipe; no Kumo entry.
TabBench-Bio: TabPFN-4 is the mean of 5 seeds; no CI.
BeyondArena: TabPFN-4 leads the Pareto front
1 / 2
TabPFN-4TabPFN-3.5Named baselinesOther baselines
Evaluation details
Baselines as of November 17, 2026. TabPFN-4 timed on an RTX PRO 6000 following the official benchmark protocol. Median inference time per 1,000 test rows. Dashed line: Pareto front.
TabPFN-4TabPFN-3.5Named baselinesOther baselines
Evaluation details
Baselines as of November 17, 2026. TabPFN-4 timed on an RTX PRO 6000 following the official benchmark protocol. Median inference time per 1,000 test rows. Dashed line: Pareto front.
TabArena: TabPFN-4 leads the Pareto front
1 / 2
TabArena: Elo and inference time
TabPFN-4TabPFN-3.5Named baselinesOther baselines
Evaluation details
Baselines as of November 17, 2026. TabPFN-4 timed on an RTX PRO 6000 following the official benchmark protocol. Median inference time per 1,000 test rows. Dashed line: Pareto front.
TabArena: improvability and inference time
TabPFN-4TabPFN-3.5Named baselinesOther baselines
Evaluation details
Baselines as of November 17, 2026. TabPFN-4 timed on an RTX PRO 6000 following the official benchmark protocol. Median inference time per 1,000 test rows. Dashed line: Pareto front.
Scaling to 10 million rows
Extrapolated: 3.5 ≈65 h, Kumo ≈127 h
1 / 2
Extrapolated: 3.5 ≈65 h, Kumo ≈127 h
TabPFN-4TabPFN-3.5Kumo-Tabular L
dashed: quadratic fit, extrapolated
Evaluation details
Synthetic data, 500 columns, 1,000 test rows, one GPU.
One estimator timed, scaled by the default count.
Measured: 38 min at 10M vs 48 min at 1M
TabPFN-4TabPFN-3.5Kumo-Tabular L
Evaluation details
Synthetic data, 500 columns, 1,000 test rows, one GPU.
One estimator timed, scaled by the default count.
Previous tabular foundation models scale quadratically with the number of rows, so large tables quickly become impractical. Extrapolated to 10 million training rows, TabPFN-3.5 would take about 65 hours and Kumo-Tabular L about 127 hours at their default estimator counts. TabPFN-4’s subquadratic architecture runs the same 10 million × 500 table in 38 minutes on a single RTX PRO 6000, roughly 100× faster than TabPFN-3.5 and 200× faster than Kumo-Tabular L. This enables fast experimentation at larger scales than any previous tabular foundation model.
Biggest gains on large, real-world data
TabPFN-4 improvements are particularly large in data regimes that are common among high-impact real-world use cases: large, temporal, and grouped data.
Until now, tuned models such as XGBoost and RealMLP, or Kumo-Tabular, were the strongest baselines in these regimes. On BeyondArena, where TabPFN-4 receives the group ID as a feature, it leads the best of them by 292 Elo on tables above 100,000 rows, 229 on temporal splits and 145 on grouped splits, compared with 42 on random splits.
Temporal and grouped data break the assumption that rows are independent draws from one distribution: the future differs from the past, and rows from the same customer or machine look alike. Classical models only handle this when someone engineers it into features. TabPFN-4’s lead is largest exactly here, which suggests a foundation model can learn to handle such structure on its own.
BeyondArena: large tables
1 / 4
BeyondArena: large tables
2,0001,8001,6001,4001,2001,000800
1,168
XGBoost
1,268
RealMLP
912
TabPFN-3
1,164
TabPFN-3.5
1,560
TabPFN-4
Elo (95% confidence intervals); bars start at 1,000
Evaluation details
BeyondArena, baseline set 1117 (official); TabPFN-4 sees the group id as a feature. XGBoost (default) = 1,000. 95% bootstrap CIs. One seed.
BeyondArena: temporal split
1,8001,6001,4001,2001,000800
1,126
XGBoost
1,173
Kumo-L
1,023
TabPFN-3
1,172
TabPFN-3.5
1,403
TabPFN-4
Elo (95% confidence intervals); bars start at 1,000
Evaluation details
BeyondArena, baseline set 1117 (official); TabPFN-4 sees the group id as a feature. XGBoost (default) = 1,000. 95% bootstrap CIs. One seed.
BeyondArena: grouped split
1,6001,4001,2001,000800
1,095
XGBoost
1,246
RealMLP
1,048
TabPFN-3
1,156
TabPFN-3.5
1,391
TabPFN-4
Elo (95% confidence intervals); bars start at 1,000
Evaluation details
BeyondArena, baseline set 1117 (official); TabPFN-4 sees the group id as a feature. XGBoost (default) = 1,000. 95% bootstrap CIs. One seed.
BeyondArena: random split
1,8001,6001,4001,2001,000
1,122
XGBoost
1,574
Kumo-L
1,333
TabPFN-3
1,496
TabPFN-3.5
1,616
TabPFN-4
Elo (95% confidence intervals); bars start at 1,000
Evaluation details
BeyondArena, baseline set 1117 (official); TabPFN-4 sees the group id as a feature. XGBoost (default) = 1,000. 95% bootstrap CIs. One seed.
The strength of TabPFN-4 in these data regimes directly leads to large improvements on benchmarks that make use of such messy, complex, real-world data. Just by switching out the model in the TabPFN-Rel harness, we see an uplift of over 100 Elo points on relational data (RelArena-alpha). On high-dimensional biomedical data (TabBench-Bio), TabPFN-4 outperforms the next-best baseline by over 50 Elo points on average.
RelArena-α: TabPFN-4 against the top relational models and systems
1 / 2
RelArena-α: TabPFN-4 against the top relational models and systems
TabPFN-4
2,055
TabPFN-Rel API (3.5)
1,914
kurversc
1,908
TabPFN-Rel local (3.5)
1,878
relflow
1,855
RT-J
1,839
RT-PluRel
1,784
1,5001,7001,9002,1002,300
Elo (95% confidence intervals); axis starts at 1,500
Evaluation details
RelArena-α public leaderboard, models and systems; newest entry per family, plus TabPFN-Rel local. Constant baseline = 1,000.
95% bootstrap CIs over tasks. One seed.
TabBench-Bio: mean Elo by task, sample budget and feature cap
TabPFN-4TabPFN-3.5Kumo-Tabular LMLPCatBoost
Evaluation details
Mean Elo over the benchmark's 28 settings. TabPFN-4 was evaluated by us with the benchmark's pipeline, all other models as published.
Tables with text and images
We evaluate TabPFN-4 on two multimodal tabular benchmarks, STRABLE and MulTaBench. Every model gets the same encoder: TF-IDF for STRABLE, e5-small-v2 for MulTaBench Text, and DINOv3-small for MulTaBench Image. TabPFN-4 ranks first on all of them.
On the text benchmarks we also show TabPFN-4 Plus, the default in our API, whose built-in text handling pushes the lead further.
STRABLE
1 / 3
STRABLE
TabPFN-4 Plus
2,010
TabPFN-4 (TF-IDF)
1,879
Kumo-Tabular (TF-IDF)
1,828
TabPFN-3.5 (TF-IDF)
1,620
1,5001,6621,8251,9882,150
Elo (95% confidence intervals); axis starts at 1,500
MulTaBench (text)
TabPFN-4 Plus
1,917
TabPFN-4
1,843
Kumo-Tabular (Frozen-S)
1,813
TabPFN-3.5 (Frozen-S)
1,637
1,5001,6621,8251,9882,150
Elo (95% confidence intervals); axis starts at 1,500
MulTaBench (images)
TabPFN-4
1,874
Kumo-Tabular (Frozen-S)
1,856
TabPFN-3.5 (Frozen-S)
1,762
1,5001,6621,8251,9882,150
Elo (95% confidence intervals); axis starts at 1,500
TabPFN-4 on Kaggle
Tabular competitions have always been among the most contested on Kaggle. Winning meant months of feature engineering, hyperparameter tuning, and deep ensembles, built against a private test set nobody got to see. Tabular foundation models have been improving rapidly over the past few months and have reached a point where they would have matched or beaten the winners of several of those competitions with a single fit.
We made late submissions with TabPFN-4 to 38 closed competitions: 25 Playground Series episodes and 13 older prize-money competitions. One forward pass per competition with default settings, no tuning, scored on the private leaderboard.
Private-leaderboard rank of one late submission per competition: one fit on the raw training data, default settings, no tuning, one run.
Zones follow Kaggle's medal cut-offs for the number of teams; late submissions earn no medals. Rank = 1 + teams with a strictly better score.
TabPFN-4 would have tied or beaten the original winner on 14 of the 38 and finished in the gold zone on 8 of the 13 prize-money competitions. On prize-money competitions, that is 7 wins, two more than Kumo-Tabular and three more than TabPFN-3.5. In the 25 Playground Series competitions, it would have taken first place on the private leaderboard in 7, ahead of every human team; Kumo-Tabular and TabPFN-3.5 would have won 2 each. Tabular foundation models have become the de facto starting model for Kagglers on tabular problems, and TabPFN-4 is now the strongest of them.
What's Next
We are working with customers to apply TabPFN-2.5 Scaling Mode to their largest datasets.